When Advice Becomes Infrastructure: Ethical Governance of Conversational AI in Psychoactive Substance Information Ecosystems

Abstract

Public debates about psychoactive substances have traditionally been organized around the pharmacology of compounds and the institutional control of supply. In digitally mediated societies, however, the pathways through which people encounter psychoactives are increasingly informational: search engines, recommender systems, social platforms, and—distinctively—conversational AI. These systems do not merely deliver neutral facts. They rank, frame, personalize, and conversationally validate claims in ways that can shape perceived norms, acceptable risk thresholds, and willingness to seek help. This opinion advances the concept of AI-mediated exposure to capture how algorithmic curation and interactive dialogue become upstream determinants of psychoactive-related harms and benefits across the continuum from everyday medicines to non-medical use. From a social-scientific ethics perspective, the central question is not whether AI is “good” or “bad,” but what obligations apply when AI performs interpretive authority in contexts characterized by vulnerability, stigma, and unequal access to trusted expertise. The paper argues for an ethics-centered governance framework grounded in four commitments: epistemic responsibility (how claims are generated, warranted, and communicated), relational responsibility (how users are treated in moments of uncertainty, distress, and stigma), distributive justice (who benefits and who bears risk under unequal conditions), and accountability (how behavior is evaluated, contested, and corrected over time). The aim is to treat conversational AI as a public-facing institution whose design choices must be ethically legible and publicly contestable, oriented toward harm reduction without intensifying surveillance, moralization, or inequity.

1. Introduction

Psychoactive substances occupy a peculiar social and political position. They are simultaneously therapeutic tools, commodities, objects of desire and fear, and symbolic markers of morality and deviance. Consequently, the harms and benefits associated with psychoactives are never determined solely by molecular properties. They are co-produced by institutions and narratives: prescribing cultures and pharmacy practices, social norms and stigma, markets and marketing, policing and punishment, and the distribution of healthcare access [1,2]. For decades, the dominant policy imagination has treated psychoactive risk as a function of exposure to substances and of the regulatory control of supply. That paradigm remains important, but it is increasingly incomplete.

A major transformation has taken place in the infrastructure of everyday knowledge. Large segments of the population now learn about psychoactives—whether medicines, alcohol, stimulants, or illicit drugs—through algorithmically curated platforms. Search engines and social media do more than provide information; they structure attention [3]. They determine which claims become salient, which communities are discoverable, which experiences are normalized, and which dangers are foregrounded or minimized. In this environment, conversational AI introduces a further shift: it does not merely present a set of sources but produces a conversational synthesis that feels tailored, coherent, and responsive [4]. It thereby assumes a role that resembles an interpreter, counselor, or advisor. This role is ethically consequential because psychoactive-related decisions are often made under conditions of uncertainty, distress, and stigma. The impression of being advised can carry more force than reading a static page, even when the content is similar.

This opinion argues that conversational AI should be understood as part of the public-facing infrastructure that shapes psychoactive substance use. The concept of infrastructure is not metaphorical. Like other infrastructures—healthcare triage, emergency hotlines, school counseling, consumer labeling—conversational systems routinize access to knowledge, influence decision-making at scale, and embed normative assumptions into everyday interactions. When a tool is regularly used to interpret psychoactive risk, it becomes an institutional actor whether or not it is formally recognized as one. The ethical challenge, therefore, is to specify what responsibilities follow from this institutional role. The paper develops the concept of AI-mediated exposure: a pathway through which algorithmic curation and interactive dialogue become upstream determinants of psychoactive-related behaviors and outcomes. The central contribution is to treat conversational AI as a public-facing institution in psychoactive information ecosystems and to specify ethically grounded governance commitments that can guide design, evaluation, and oversight. The unifying thread is that conversational AI functions as an advice infrastructure: by shaping how uncertainty, normality, and legitimacy are communicated, it can shift psychoactive-related trajectories through changes in perceived norms, risk thresholds, and pathways to help-seeking. The sections that follow therefore move from this infrastructural mechanism (AI-mediated exposure and interpretive authority) to a governance response (four commitments) that specifies what responsible design, evaluation, and oversight require in a stigma- and inequality-shaped domain. It then proposes an ethics-centered governance framework for conversational AI in psychoactive contexts, oriented toward harm reduction, dignity, and justice rather than punitive drift or technological solutionism.

In this paper, conversational AI is used as a shorthand for a heterogeneous set of systems rather than a single, uniform actor. Commercial products, academic prototypes, and open-source models differ in training data, safety policies, interface design, and deployment incentives, which can yield materially different patterns of refusal, reassurance, and redirection. Moreover, performance and user experience can vary across languages, cultural contexts, and levels of health literacy, shaping who can effectively interpret uncertainty, follow harm-reduction guidance, or access care pathways. Accordingly, the ethical obligations discussed here should be understood as governance expectations for high-trust conversational interfaces in psychoactive information contexts, while recognizing that concrete risks and mitigations are system- and population-specific. This framework is offered as an ethical and governance lens for conversational AI as a public-facing information infrastructure; it is not a substitute for clinical judgment, pharmacological expertise, or regulatory decision-making. Rather, it aims to complement these domains by clarifying the ethical obligations and institutional design constraints that arise when AI performs interpretive authority under uncertainty, stigma, and unequal access to care.

2. From Molecules to Information Pathways: AI-Mediated Exposure

Exposure has typically been imagined as physical contact with a substance. In digitally mediated societies, many psychoactive trajectories begin earlier—with contact with claims. These claims include narratives of relief and transformation, warnings about dangers, practical advice, peer stories, and cultural scripts about what responsible use looks like. The information ecosystem shapes whether individuals initiate use, escalate, mix substances, discontinue abruptly, conceal behaviors, or seek help. This is not a peripheral issue. For many people, a point of contact with psychoactive-related knowledge is not a clinician or pharmacist but a platform [5].

Conversational AI intensifies the ethical significance of information pathways for three reasons. First, it produces a coherent narrative, often smoothing over uncertainty and conflict. Second, it responds to the user’s expressed needs, creating a sense of recognition and personalization. Third, it performs a relational stance—empathetic, neutral, or authoritative—which can influence trust and compliance. Even when the system intends to be cautious, the conversational form itself can create the impression of counsel. A social-scientific ethics lens treats this not as a narrow content-moderation issue but as a question of power: who gets to frame what counts as safe, normal, or legitimate; whose experiences are amplified; and how uncertainty is communicated. These are not merely technical design choices. They are institutional choices with distributive consequences. Importantly, AI-mediated exposure operates across the psychoactive continuum [6]. The distinction between medicine use and drug use is socially and politically loaded, but informational pathways often blur it [7]. People may search for a medicine to manage symptoms, encounter discussions of non-medical uses, move into communities where risk is reframed as experimentation, or conversely encounter harm-reduction resources that redirect behavior toward safer practices and support. An ethical framework must therefore address how conversational systems shape trajectories across the continuum rather than assuming a stable category of user. This study uses “psychoactive substances” in an intentionally inclusive sense, spanning prescribed medications, alcohol and other legal substances, and illicit drugs. At the same time, the ethical stakes are not uniform across this spectrum: legal risk, stigma, typical user vulnerability, and the appropriateness of clinical referral can differ dramatically by substance category and use context. The continuum framing is meant to highlight shared informational pathways and governance responsibilities, not to collapse these important differences.

3. Conversational Authority and the Ethics of Plausible Counsel

A defining feature of conversational AI is the performance of interpretive authority. The system’s responses can sound confident, organized, and responsive even when the underlying evidence is incomplete or context-dependent [8]. This produces a distinctive ethical risk: the substitution of plausibility for warranted knowledge. While misinformation is an obvious concern, the deeper issue is epistemic overreach—the appearance of competence where competence is not established. In psychoactive contexts, epistemic overreach can lead to two broad harms. The first is behavioral: users may treat conversational output as sufficient grounds for decisions that require professional assessment or careful contextual judgment. The second is normative: the system’s framing may normalize certain behaviors, downplay structural constraints, or implicitly moralize users’ concerns. The latter matters because psychoactive use is deeply entangled with stigma. A system that communicates blame, suspicion, or punitive assumptions can increase concealment and reduce help-seeking, even if it avoids explicit errors. Illustratively, a user asking whether to stop an antidepressant “for a few days” may interpret a fluent, confident reply as permission to discontinue, whereas calibrated uncertainty and tapering-oriented redirection can shift the decision toward safer follow-up. Similarly, when asked whether combining a sedative with alcohol is “usually fine,” vague normalization can raise perceived acceptability, while harm-reduction framing can clarify interaction risk and suggest nonjudgmental next steps.

Classical toxicological principles underscore why this is not a generic information problem. Psychoactive risks often hinge on dose–response relationships, narrow safety margins (or therapeutic windows), and substantial interindividual variability in metabolism, tolerance, comorbidities, and co-medication profiles. When conversational AI presents generalized guidance with an aura of specificity, it can obscure these fundamentals—inviting users to treat “what is plausible on average” as safe for their particular body and context. In psychoactive contexts, epistemic overreach can translate into pharmacologically consequential actions—such as self-directed dose changes, abrupt medication discontinuation without tapering, risky combinations with interaction or toxicity potential, or the minimization of cumulative and delayed harms when short-term relief is framed as sufficient evidence of safety. Making these pathways explicit underscores that the ethical risk is not only misinformation but the conversational production of “plausible counsel” that can shift real-world exposure and toxicity trajectories.

Epistemic responsibility in conversational AI must therefore be understood as an ethical duty to communicate the limits of knowledge and the conditionality of guidance. The aim is not to overwhelm users with disclaimers but to avoid rhetorical closure—the conversational dynamic in which a user leaves feeling they have received definitive instruction in a domain that remains uncertain. Ethically responsible systems should cultivate calibrated confidence: conveying what is known, what is uncertain, what depends on context, and what requires external support. A further ethical complication arises because conversational AI often merges multiple registers—informational, emotional, and directive [9]. A response that is emotionally validating may be interpreted as endorsing a plan of action. Conversely, a refusal or caution can be interpreted as judgment. The ethics of plausible counsel, then, is not simply about factual content; it is about how conversational form shapes perceived legitimacy and permission.

4. Recommender Systems, Norm Formation, and the Politics of Visibility

While conversational AI is the focus of this paper, it exists within a broader algorithmic ecosystem. Recommender systems and platform dynamics shape what questions users ask and what narratives they find credible. Psychoactive-related content is often engagement-rich: it promises relief, transformation, rebellion, or insight, and it frequently circulates through compelling personal stories. Engagement-based ranking can unintentionally privilege emotionally salient narratives over nuanced risk communication. The ethical stakes here are about norm formation. Platforms can create the impression that a behavior is common or socially accepted through repeated exposure. They can also concentrate attention on extreme experiences, either glamorizing risk or amplifying fear. In both cases, the information ecosystem shapes perceived normality. This matters because perceived normality is a strong predictor of behavior, especially among younger users and among those seeking belonging in stigmatized domains. Conversational AI can amplify these platform effects by synthesizing and reframing them. If a user arrives after exposure to a certain narrative—everyone is doing this, this is natural and safe, professionals cannot be trusted—the conversational system’s stance becomes pivotal [10]. It may correct misperceptions, reinforce them, or redirect attention toward safer pathways. Thus, the ethical evaluation of conversational AI cannot be isolated from the political economy of platforms: what is rewarded, what is visible, and which stakeholders shape the informational environment. From a governance perspective, this implies that ethical responsibility is distributed. It belongs not only to the developers of conversational systems but also to platform operators, advertisers, content creators, and regulators. Nonetheless, conversational AI deserves special attention because it occupies a position of interpretive synthesis and can become a high-trust interface for users navigating uncertainty.

5. Vulnerability, Stigma, and Relational Responsibility in Dialogue

Many psychoactive-related queries occur under conditions of vulnerability: distress, anxiety, isolation, impulsivity, or fear of judgment [11]. People may seek information because they lack access to trusted professionals, because they fear stigma, or because they want anonymity. In such contexts, conversational AI can become emotionally salient. It can feel safe precisely because it appears nonjudgmental and readily available. This argument also connects to adjacent studies on digital and AI-mediated health advice in substance-use and addiction-related contexts, which emphasize that advice interfaces can modulate behavior not only through factual content but through trust, stigma management, and perceived access to care. Even when evidence remains emergent, these strands help situate conversational AI as part of a broader shift in how people calibrate risk, interpret symptoms, and decide whether to disclose, persist, or seek support in stigmatized domains. This relational quality creates ethical duties that go beyond content accuracy. Relational responsibility involves how the system treats users as moral subjects. Does it presume criminality? Does it moralize? Does it communicate respect and dignity? Does it reduce stigma or inadvertently reinforce it? These questions are critical as stigma is not merely an attitude; it is a mechanism of harm. Stigma increases concealment, delays care, and intensifies marginalization. An AI system that amplifies stigma may increase harm even while providing correct information. Relational responsibility also includes recognizing that users may be testing boundaries or seeking validation. In psychoactive contexts, a user may ask questions that signal ambivalence, curiosity, or self-justification. The system’s response can function as a form of social feedback. It can either gently redirect toward safer pathways or lend legitimacy to risky trajectories. The ethical challenge is to avoid both paternalistic control and permissive normalization. A social-scientific approach emphasizes that the system’s voice is not neutral. It participates in moral economy: it can enact compassion or suspicion, dignity or shame, solidarity or abandonment. Designing conversational AI for psychoactive contexts should therefore prioritize interactional ethics—how responses shape the user’s sense of being judged, supported, or coerced.

6. The Surveillance Temptation and the Need for Trust-Preserving Governance

Psychoactive governance has long been shaped by punitive logics. Surveillance and punishment are often presented as necessary to deter harm, yet they can produce counterproductive outcomes: users avoid disclosing problems, communities retreat to less safe channels, and marginalized groups bear disproportionate burdens. Conversational AI introduces new possibilities for surveillance because it can generate data-rich interactions at scale. The ethical risk is that health-oriented support becomes entangled with enforcement-oriented objectives. Even the perception of such entanglement can erode trust. If users fear that asking questions might expose them to punishment, they will avoid safer information pathways and seek guidance from less accountable sources. This dynamic is well documented in other domains where punitive governance undermines care-seeking behavior. In psychoactive contexts, it is particularly acute because stigma and legal risk already discourage disclosure. Trust-preserving governance requires a bright line between public health support and punitive surveillance. This does not entail ignoring legality; rather, it means that the primary design objective for psychoactive-related AI interactions should be harm reduction and care facilitation, not the production of traceable evidence [12]. Ethical governance can minimize data retention, limit secondary use, and adopt purpose limitation principles that restrict how interaction data can be used. The legitimacy of conversational AI as a public-facing institution depends on whether users can reasonably expect that seeking information will not expose them to disproportionate punitive risk. A social-scientific ethics perspective also insists that surveillance is not evenly distributed. Those already marginalized—by poverty, homelessness, incarceration exposure, or racialized policing—face higher risks from data misuse and enforcement drift. Therefore, governance that ignores these structural asymmetries is ethically inadequate even if it appears neutral.

7. Distributive Justice and the Unequal Burden of AI-Mediated Risk

Conversational AI enters a landscape of unequal access to expertise [13]. Individuals with stable healthcare access can cross-check AI outputs with clinicians and pharmacists. Those without such access may rely on AI as a primary advisor, especially for stigmatized concerns. This creates a distributive justice problem: uncertainty, error, and persuasive misframing will be borne disproportionately by those with fewer resources to verify and correct. Justice concerns also arise in how systems allocate attention and legitimacy. If conversational AI consistently emphasizes individual responsibility while minimizing structural drivers of psychoactive-related harm—economic precarity, trauma, social isolation, inadequate care access—it risks reproducing moral narratives that blame individuals for problems shaped by institutions. Conversely, if it overemphasizes structural explanations without offering pathways to agency and support, it can produce fatalism. Ethical design must navigate this tension by acknowledging structural constraints while supporting practical, dignity-preserving steps toward safety and help.

Further, distributive justice includes accessibility. If conversational systems use language that presumes high literacy, medical familiarity, or cultural alignment with dominant norms, they will serve some users better than others [14,15]. In psychoactive contexts, inequitable comprehension can translate into inequitable risk. Justice-oriented governance therefore requires that conversational AI be evaluated for disparate impacts: who benefits, who is harmed, who is excluded, and whose needs are invisibilized. Finally, the commercialization of AI services can exacerbate inequity. If high-quality guidance is placed behind paywalls or if systems steer users toward paid services as default solutions, then those with fewer resources may receive less safe or less supportive interactions. Ethical governance must therefore consider the political economy of access, not only the content of responses.

8. Accountability as Public Contestability: Governing AI as an Institution

Infrastructures that shape public behavior require accountability mechanisms that go beyond voluntary commitments. For conversational AI in psychoactive contexts, accountability must be understood as public contestability: the capacity of affected stakeholders to scrutinize system behavior, challenge harmful patterns, and demand changes. Public contestability has several dimensions. Meaningful public contestability can be enabled through a small set of institutional mechanisms without exposing sensitive user data. Platforms can (i) publish accessible governance notes describing high-level response policies and major changes over time; (ii) provide structured channels for clinicians, harm-reduction groups, and affected communities to report recurring harmful patterns and receive documented remediation; and (iii) support independent, privacy-preserving audits using standardized test scenarios, with aggregate results reported across languages and health-literacy conditions. These mechanisms make behavioral patterns observable and correctable while preserving trust and privacy. First, system behavior must be ethically legible. Users and stakeholders should be able to understand why certain kinds of responses occur—whether a system tends to overconfidently reassure, to moralize, to refuse, or to redirect. Second, there must be pathways for contestation and correction. If a community observes stigmatizing patterns or harmful refusal dynamics, there should be structured channels for remediation. Third, accountability must be ongoing. Conversational AI systems change over time; cultural narratives shift; new psychoactive trends emerge. One-time audits are insufficient. Accountability is also about what is measured. Many platform governance efforts focus on the presence or absence of prohibited content. In psychoactive contexts, however, ethical evaluation must examine downstream social effects: whether system behavior discourages help-seeking, increases shame, contributes to normalization of risk, or disproportionately harms marginalized users. These are not easily captured by simple metrics, but they are central to ethical legitimacy. A governance framework grounded in public contestability resists two failures that frequently distort debates. It resists the censorship reflex that treats the problem as purely the elimination of content. And it resists technological solutionism that treats AI as a substitute for care. Instead, it treats conversational AI as a public-facing institution that must be governed through transparent norms, continuous evaluation, and mechanisms for democratic oversight.

9. An Ethics-Centered Governance Framework for Conversational AI in Psychoactive Contexts

This section presents the paper’s core governance contribution: a four-commitment framework for ethically evaluating and guiding conversational AI in psychoactive contexts. Analytically, this ethics-centered framework is organized around four commitments: epistemic responsibility, relational responsibility, distributive justice, and accountability. These commitments are not intended to imply that ethical governance will always be tension-free in practice. In real-world cases, harm reduction, respect for autonomy, stigma-sensitive communication, and institutional accountability may not always align perfectly and may require contextual interpretation and reasoned balancing. For that reason, the framework is not offered as a one-size-fits-all formula or a rigid moral blueprint, but as a structured basis for context-sensitive judgment in psychoactive-related information environments. In this sense, the four commitments are best understood as guiding considerations for system design, evaluation, and institutional oversight rather than mechanically applicable rules. To make these commitments more operational, the following discussion offers brief applied implications for design, auditing, and policy. For relational responsibility, examples include stigma-reducing language, nonjudgmental refusals, supportive check-ins when users signal distress, and harm-reduction redirection that preserves dignity. For distributive justice, practical evaluation can include disparity audits across languages and health-literacy conditions, accessibility testing (reading level/jargon), and analysis of referral patterns (public vs. commercial resources; local vs. generic pathways). These examples are illustrative rather than exhaustive, intended to translate high-level commitments into actionable governance expectations. Epistemic responsibility requires that conversational AI communicate uncertainty honestly, distinguish general information from individualized counsel, and avoid rhetorical closure. The ethical goal is to prevent the performance of unwarranted authority. This includes attention to tone and certainty, since the persuasive force of conversation can create compliance even when the system’s knowledge is limited. Relational responsibility requires that the system’s interactional behavior reduce stigma and support dignity.

Applied implications for design, auditing, and policy can be illustrated with brief examples. For epistemic responsibility, systems can implement confidence calibration (e.g., explicitly marking what is uncertain or context-dependent), interaction-risk prompts (e.g., “mixing substances can be dangerous; if you share what you took and when, I can give general safety-oriented information”), and clear boundaries between general information and individualized counsel. For relational responsibility, interaction design can prioritize stigma-reducing language, nonjudgmental refusals (avoiding moralized or criminalizing tone), supportive check-ins when users signal distress or shame, and harm-reduction redirection that preserves dignity (e.g., offering locally actionable support options without coercive fear). For distributive justice, concrete safeguards include plain-language modes, multilingual parity goals, and routing to non-paywalled public resources alongside any commercial options, with monitoring to prevent systematic steering that disadvantages low-resource users. For accountability, institutions can publish transparent governance notes (what the system tends to refuse, reassure, or redirect), enable structured user/community feedback channels, and support independent audits focused on downstream effects (help-seeking, stigma, normalization) rather than only prohibited-content counts.

A brief contrast illustrates how the same commitments can apply differently across contexts. In a query about antidepressant side effects or discontinuation, epistemic responsibility emphasizes calibrated discussion of uncertainty and individualized risk, while relational responsibility prioritizes stigma-reducing reassurance that encourages appropriate clinical follow-up (e.g., avoiding abrupt cessation and supporting safe help-seeking). In a query about recreational psychedelic use, the legal and situational context heightens the importance of nonjudgmental harm-reduction framing, careful avoidance of overconfident “how-to” counsel, and trust-preserving redirection to credible safety resources—while also recognizing that users may avoid formal care due to stigma or legal fear. Across both cases, distributive justice and accountability require attention to who can verify guidance, how refusals or redirections are experienced, and whether the system’s behavior disproportionately burdens marginalized users.

The system should avoid construing psychoactive-related vulnerability, distress, or help-seeking as signs of moral failure or presumptive criminality. However, this relational commitment should not be read as suspending ethical or legal judgment in cases involving concrete harm to others; where coercion, exploitation, neglect, abuse, or serious safeguarding concerns arise, supportive design must remain compatible with accountability. Relational responsibility also means treating vulnerability as an ethical condition: when users disclose fear, shame, or crisis signals, the system should respond in ways that orient toward safety and support rather than merely maintaining conversational fluency. Distributive justice requires that governance anticipate unequal capacity to verify information and unequal exposure to punitive consequences. Systems can be evaluated for disparate impacts and should not amplify inequity through inaccessible language, commercially biased pathways, or the normalization of narratives that blame individuals for structurally produced harm. Accountability requires public contestability and continuous evaluation. Because conversational AI is dynamic and context-dependent, ethical governance must be iterative. It must allow for scrutiny, revision, and community-informed correction. Accountability also requires resisting surveillance drift by adopting trust-preserving data practices and ensuring that support functions are not repurposed into punitive instruments. This framework is intended to guide how stakeholders should think about governance rather than to offer a single policy solution. It clarifies what is at stake ethically when conversational AI becomes part of the psychoactive information ecosystem: the shaping of norms, the distribution of risk, the protection of dignity, and the maintenance of trust.

10. Evaluation as an Ethical Practice

Evaluation is not merely technical; it is ethical because it defines which harms count and whose experiences matter. In psychoactive contexts, evaluating conversational AI solely by whether it blocks certain content categories misses the broader institutional effects. Ethical evaluation should ask whether the system’s behavior supports safer trajectories, reduces stigma, and facilitates help-seeking without generating coercive fear or punitive exposure. In practice, the framework can be operationalized through observable behaviors and lightweight audits. Epistemic responsibility can be tested with scenario-based prompts that check uncertainty communication and avoidance of overconfident closure in time-sensitive or interaction-risk queries. Relational responsibility can be assessed by coding for stigma-reducing language, nonjudgmental refusals, and supportive redirection to care. Distributive justice can be evaluated via disparity and accessibility testing across languages and health-literacy conditions and by monitoring referral patterns (public vs. commercial; local vs. generic resources). Accountability can be supported through governance documentation (versioning/policy-change logs) and structured stakeholder feedback channels with tracked remediation.

As the harms of infrastructure are often diffuse and cumulative, evaluation should attend to patterns rather than isolated failures. Does the system consistently over-reassure? Does it disproportionately refuse in ways that shame users? Does it treat certain populations as suspicious? Does it route users toward commercial services rather than public resources? Does it inadvertently reproduce stereotypes about who uses psychoactives and why? These questions require methods that combine qualitative and quantitative assessment, including attention to user experience, community feedback, and equity impacts. Continuity is especially salient in rapidly evolving pharmacological domains, where evidence, guidelines, and risk signals can change quickly. Conversational systems may not reliably reflect recent updates due to model versioning, uneven knowledge refresh cycles, or differences between curated clinical guidance and internet-salient narratives. Ethical evaluation should therefore document the system version and update practices, test time-sensitive queries, and examine whether outdated or overconfident outputs are disproportionately produced in non-dominant languages or among users with lower health literacy.

Ethical evaluation must also be continuous. In practice, continuous evaluation can be implemented through lightweight test suites and periodic audits that track recurring patterns (e.g., over-reassurance, moralized refusal, or weak uncertainty communication). Results should be reported in aggregate, stratified by language and health-literacy conditions, and complemented by referral-pattern monitoring to ensure users are routed to credible, locally actionable, non-paywalled resources when appropriate. Conversational AI systems change; platform cultures shift; psychoactive narratives evolve. A system that appears responsible in one moment can become harmful as contexts change. Ongoing stewardship—monitoring, learning, and revising—is therefore an ethical obligation for institutions operating conversational AI at scale. Evaluation must remain compatible with trust. Collecting data to improve safety should not become a pretext for surveillance. Ethical evaluation should emphasize privacy-preserving approaches, aggregate analysis where possible, and clear limits on secondary use. The legitimacy of conversational AI in psychoactive contexts depends on whether evaluation practices align with harm reduction and dignity rather than enforcement and punishment.

In practice, several lightweight methodologies can operationalize these commitments. Disparity audits can test whether safety-relevant guidance differs systematically across languages, cultural contexts, and health literacy levels (e.g., whether uncertainty is communicated less clearly or refusals become more moralizing in non-dominant languages). Accessibility testing can assess reading level, jargon density, and whether key risk messages (e.g., interactions, tapering, delayed harms) remain comprehensible under simplified literacy conditions. Referral-pattern analysis can evaluate distributive justice by examining where users are routed: public versus commercial resources, local versus generic options, and whether stigmatized users receive fewer actionable pathways to care. Finally, behavioral outcome proxies—such as changes in stated willingness to seek help, perceived stigma, or perceived normality after an interaction—can be measured in controlled studies or structured user testing to detect systematic over-reassurance, shame-inducing refusals, or normalization of risk.

11. Conclusions

Conversational AI is becoming an institutional actor in psychoactive substance information ecosystems, not because it is formally designated as such, but because it increasingly performs functions that institutions traditionally held: interpreting uncertainty, shaping norms of reasonable conduct, and providing language that users experience as guidance. Its ethical significance therefore lies not only in the truth-conditional content of any single response but in the cumulative effects of repeated interactions at scale. When many micro-encounters are mediated through a system that speaks fluently, empathically, and with apparent coherence, the system participates in producing what counts as credible knowledge, what feels socially acceptable, and which actions appear legitimate in moments of vulnerability.

A central implication is that psychoactive-related governance cannot remain molecule-centered while the primary pathways of meaning-making become platform-centered. As conversational systems synthesize and reframe narratives users have already encountered through algorithmic feeds, they can amplify platform effects by consolidating a storyline into a personalized takeaway. In psychoactive contexts, where stigma, fear of judgment, and ambivalence are common, the system’s stance becomes pivotal. It can normalize risky narratives, subtly moralize and deter disclosure, or reorient the interaction toward dignity-preserving safety and support. This is why the ethical challenge is not reducible to preventing explicit errors. It concerns how interpretive authority is performed, how uncertainty is communicated, how vulnerability is handled, and how the interaction constructs the user as either a legitimate help-seeker or a suspect subject.

Ethical governance must resist two familiar forms of drift. The first is punitive drift, where support infrastructures become instruments of surveillance. In psychoactive contexts, the mere perception that conversational interactions might generate traceable evidence can chill help-seeking and push users toward less accountable channels. Trust is not an accessory here. It is a precondition for harm reduction. The second is solutionist drift, where conversational AI is treated as a substitute for institutions of care rather than as a mediator within an unequal care landscape. If AI systems become default advisors for those with the least access to professional expertise, then shortcomings such as overconfidence, inaccessible language, and culturally narrow norms will be distributed inequitably. Infrastructures that are good enough for privileged users can still deepen harm for those who cannot verify, contest, or navigate them.

For these reasons, an ethics-centered approach should treat conversational AI as a public-facing institution whose legitimacy depends on epistemic responsibility, relational responsibility, distributive justice, and accountability. Epistemic responsibility requires calibrated confidence and the avoidance of rhetorical closure in domains where context is decisive. Relational responsibility requires conversational design that reduces stigma, preserves dignity, and supports disclosure without coercion. Distributive justice requires attention to literacy, culture, and unequal access to expertise so that conversational systems do not quietly privilege those already advantaged. Accountability requires public contestability and continuous evaluation, because the power to frame psychoactive risk and normality is not merely technical power but social power.

Ultimately, the question is not whether conversational AI should participate in psychoactive-related information pathways, because it already does. The question is what kind of institution it will become. If governed primarily through market incentives and risk-avoidant compliance, it may default to patterns that amplify harmful narratives or deter vulnerable users through moralized refusal and surveillance-like features. If governed as a trust-preserving public health infrastructure, it can support safer trajectories by offering nonjudgmental orientation, clarifying uncertainty, and facilitating pathways to care without punitive exposure. Recognizing conversational AI as infrastructure makes its ethical stakes visible. It is a site where harm reduction, stigma, justice, and civic legitimacy will be negotiated in everyday language, one interaction at a time.

Practical implications follow for pharmacology, toxicology, and public health. For pharmacology and toxicology, high-trust conversational interfaces should be evaluated for how they communicate dose–response sensitivity, safety margins/therapeutic windows, tapering and withdrawal dynamics, and interaction/toxicity risks from substance combinations, including cumulative or delayed harms. For public health, systems can be assessed for whether their conversational stance reduces stigma, supports timely help-seeking, and routes users toward credible, locally actionable resources without coercive fear or surveillance drift. Across domains, the core governance task is to make uncertainty and context-dependence legible, prioritize harm-reduction pathways, and ensure that accountability mechanisms can detect and correct systematic patterns of over-reassurance, moralized refusal, or inequitable performance across languages and health literacy levels.

12. Future Directions

Although the present paper is primarily normative and conceptual, several of the empirical implications and evaluative questions that follow from its framework can be investigated systematically. Rather than treating the ethical commitments themselves as directly testable hypotheses, future research can examine how conversational form, confidence calibration, relational stance, and norm framing shape user perceptions, trust, perceived normality, willingness to seek help, and responses to uncertainty in psychoactive-related contexts.